CN109344893B - Image classification method based on mobile terminal - Google Patents

Image classification method based on mobile terminal Download PDF

Info

Publication number
CN109344893B
CN109344893B CN201811119618.4A CN201811119618A CN109344893B CN 109344893 B CN109344893 B CN 109344893B CN 201811119618 A CN201811119618 A CN 201811119618A CN 109344893 B CN109344893 B CN 109344893B
Authority
CN
China
Prior art keywords
neural network
deep neural
weight
codebook
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201811119618.4A
Other languages
Chinese (zh)
Other versions
CN109344893A (en
Inventor
陈靓影
徐如意
饶川
刘乐元
张坤
彭世新
刘小迪
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Central China Normal University
Original Assignee
Central China Normal University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Central China Normal University filed Critical Central China Normal University
Priority to CN201811119618.4A priority Critical patent/CN109344893B/en
Publication of CN109344893A publication Critical patent/CN109344893A/en
Application granted granted Critical
Publication of CN109344893B publication Critical patent/CN109344893B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses an image classification method based on a mobile terminal. The weight in the deep neural network model is quantized into an index of 2 in training, so that the shift operation of the embedded system can be effectively improved, the quantization codebook is dynamically updated, the quantization error can be effectively reduced, and the prediction performance of the model and the operation efficiency of the model in a moving state are improved. The invention also provides a system for realizing the method. The deep neural network compression method has an obvious compression effect on the deep neural network, can reduce the consumption of a large deep neural network model on storage resources and calculation resources, promotes the deployment of the deep neural network on resource-limited mobile terminals such as smart phones and the like, and has extremely strong practical application.

Description

Image classification method based on mobile terminal
Technical Field
The invention belongs to the field of image processing and pattern recognition, and particularly relates to an image classification method and system based on a mobile terminal.
Background
With the rapid development of internet technology, the realization of high-pixel photographing function in smart phones and the universal coverage of mobile communication networks make a large amount of image information rush into our lives. Therefore, how to distinguish massive image data into different categories on a mobile terminal such as a smart phone is a technical problem which needs to be solved urgently.
In recent years, the deep neural network is distinguished from a plurality of machine learning methods, and the performance of image classification is remarkably broken through, so that the deep neural network attracts wide attention. In order to obtain better characteristics and improve the performance of the neural network, a deep multi-layer network structure is often constructed. This results in deep neural networks with millions of parameters, consuming a significant amount of computing and memory resources. And the application of the deep neural network to mobile terminals such as smart phones and the like is greatly difficult.
In order to deploy a deep neural network to embedded devices such as a smart phone, a currently common method is to compress a depth model, so that the classification performance is ensured, and meanwhile, the storage space required by the depth model is reduced as much as possible. At present, many researches have been carried out in the field, but the problems of difficult convergence of compression network training, low classification precision, low operation efficiency in a mobile terminal and the like still exist.
Disclosure of Invention
Aiming at the problems and improvement requirements in the prior art, the invention provides an image classification method and system based on a mobile terminal, which quantizes the weight in a deep neural network model into an index of 2 and can effectively improve the shift operation of an embedded system. Different from the existing method adopting static quantization coding, the method provided by the invention can be used for dynamically updating the quantization codebook in the model training process, effectively reducing the quantization error and improving the prediction performance of the model and the operation efficiency in the moving state.
An image classification method based on a mobile terminal comprises an off-line training stage and an on-line classification stage:
the off-line training stage specifically comprises:
s1 formulating a codebook:
acquiring the maximum value of the weight absolute value in each layer of the deep neural network model by adopting a deep neural network model, and quantizing the maximum value into an exponential form with the base number of 2, thereby obtaining the quantization upper limit of the codebook; determining a codebook quantized by a current model under the limit of an upper quantization limit;
s2 quantization weight:
carrying out exponential quantization on the weight in the deep neural network model, and quantizing the weight into a value closest to a codebook;
s3 retraining the network model:
inputting a sample image, training the quantized deep neural network model, obtaining the cross entropy loss of the deep network in the forward process of training, and updating the weight parameters in the network according to the cross entropy loss in the backward process.
S4 iteration and termination:
iteratively executing steps S2 and S3 until the deep neural network model converges or reaches the set training times, terminating the iteration and obtaining a final classifier;
the online classification stage specifically comprises: and sending the image to be classified into a classifier to obtain a classification result.
Further, the quantization upper limit of the codebook is expressed as:
Figure BDA0001810470700000031
which is composed of n2＝floor(log2(max(|WlI))) floor (. cndot.) is a downward rounding operation, wlFor the l-th layer weight of the deep neural network, max () represents the maximum value, and | represents the absolute value.
Further, when quantized to b bits, the codebook is expressed as: pl＝{±2n}，n∈[n1，n2]N ∈ Z, where l represents the l-th layer of the deep neural network, n1And n2Is two integers satisfying n1＜n2，n1＝n2-2b-1+1, Z represents a positive integer.
Further, when quantized to b bits, the codebook is expressed as:
Pl＝{±2n，0}，n∈[n1，n2]n ∈ Z, where l represents the l-th layer of the deep neural network, n1And n2Is two integers satisfying n1＜n2，n1＝n2-2b-2+1, Z represents a positive integer.
Further, the layers of the deep neural network are quantified as follows:
Figure BDA0001810470700000032
wherein the content of the first and second substances,
Figure BDA0001810470700000033
for quantized weights, 2kA quantized value that is the absolute value | w | of the weight w;
Figure BDA0001810470700000034
indicating function W ∈ Wl。
Further, the layers of the deep neural network are quantified as follows:
Figure BDA0001810470700000035
wherein the content of the first and second substances,
Figure BDA0001810470700000041
for quantized weights, 2kA quantized value that is the absolute value | w | of the weight w;
Figure BDA0001810470700000042
indicating function W ∈ Wl。
Further, the cross entropy loss of the deep network is obtained in the forward process of training, and the cross entropy loss is expressed as:
Figure BDA0001810470700000043
wherein the content of the first and second substances,
Figure BDA0001810470700000044
is the loss of the network and is,
Figure BDA0001810470700000045
is a regularization term, adopts L2Norm regularization term, λ is the coefficient of the regularization term,
Figure BDA0001810470700000046
is the network weight after the model compression,
Figure BDA0001810470700000047
is the total loss function.
Further, the weight parameters in the network are updated in the reverse process according to the cross entropy loss:
Figure BDA0001810470700000048
wherein the content of the first and second substances,
Figure BDA00018104707000000410
is the weight of the network at the kth iteration, gamma is the learning rate,
Figure BDA0001810470700000049
is the gradient of the loss function over the network weights.
An image classification system based on a mobile terminal comprises an offline training module and an online classification module:
the offline training phase is to:
s1 formulating a codebook:
acquiring the maximum value of the weight absolute value in each layer of the deep neural network model by adopting a deep neural network model, and quantizing the maximum value into an exponential form with the base number of 2, thereby obtaining the quantization upper limit of the codebook; determining a codebook quantized by a current model under the limit of an upper quantization limit;
s2 quantization weight:
carrying out exponential quantization on the weight in the deep neural network model, and quantizing the weight into a value closest to a codebook;
s3 retraining the network model:
inputting a sample image, training the quantized deep neural network model, obtaining the cross entropy loss of the deep network in the forward process of training, and updating the weight parameters in the network according to the cross entropy loss in the backward process.
S4 iteration and termination:
iteratively executing steps S2 and S3, and when the deep neural network model converges or reaches the set training times, terminating iteration to obtain a final classifier;
the online classification phase is to: and sending the image to be classified into a classifier to obtain a classification result.
Compared with the prior art, the invention has the advantages and effects that:
1. the invention provides a method for dynamically updating weight parameters with larger absolute values in a codebook self-adaptive network, and reduces the influence of the quantization of the parameters on the model precision as much as possible;
2. the invention provides an alternative iteration algorithm for model training, so that weight parameters and a codebook are alternately updated, and the convergence speed of the training process is higher.
Drawings
FIG. 1 is a flowchart illustrating an implementation of an image classification method based on a mobile terminal according to the present invention;
FIG. 2 is a diagram illustrating the quantization rules of network weights
Fig. 3 is a block diagram of image classification based on a mobile terminal according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
FIG. 1 is a flow chart of an implementation of the image classification method based on the mobile terminal according to the present invention. The method comprises two stages of off-line training and on-line classification:
the off-line training stage specifically comprises:
s1 formulating a codebook: and acquiring the maximum value of the weight absolute value in each layer of the randomly initialized or pre-trained deep neural network model so as to determine the quantization upper limit of the codebook, and acquiring the codebook quantized by the current model according to the quantization digit.
In step S1, the maximum value of the absolute value of the weight in each layer of the initial unquantized depth neural network model is obtained and quantized into an exponential form with a base number of 2,the concrete quantification is shown as follows:
Figure BDA0001810470700000061
wherein n is2＝floor(log2(max(|WlI))) floor (. cndot.) is a downward rounding operation, wlIs the l-th layer weight of the deep neural network, and the l represents the absolute value. And quantizing the maximum value of the absolute value of the weight of each layer of the deep neural network to obtain the upper limit of the codebook.
In step S1, when the quantization is b bits, the codebook may be expressed as: pl＝{±2n}，n∈[n1，n2]N ∈ Z, where l represents the l-th layer of the deep neural network, n1And n2Is two integers satisfying n1＜n2And Z represents a positive integer. Since n is1And n2Between is provided with n2-n1+1 integers and the positive and negative integers in the codebook are equal in number, so that the total value in the codebook is 2 (n)2-n1+1)＝2bI.e. with n1＝n2-2b-1+1, thereby determining the codebook Pl。
In step S1, 0 may also be introduced as a quantization value into the formulated codebook, which may be denoted as Pl＝{±2n，0}，n∈[n1，n2]N is equal to Z. Since 0 cannot represent the power of 2 to the n (n is an integer), an additional bit is required to represent the quantized value of 0, where n is an integer1＝n2- 2b-2+1, other processing is unchanged.
S2 quantization weight: and quantizing the weights in the deep neural network model according to the codebook established in the step S1, and quantizing the weights into the closest value in the codebook.
In step S2, the weights are quantized to quantize the weights in the deep neural network to the value nearest to the codebook, and the specific quantization rule is as shown in fig. 2, and the quantization rules for the layers of the deep neural network are as follows:
Figure BDA0001810470700000071
wherein the content of the first and second substances,
Figure BDA0001810470700000072
for quantized weights, 2kA quantized value that is the absolute value | w | of the weight w;
Figure BDA0001810470700000073
w∈Wlis an indication function for distinguishing positive and negative weights in the network.
In step S2, 0 is introduced into the codebook as a quantization value, and the lower limit of the codebook is truncated to obtain a 0 value, and the corresponding quantization method is as follows:
Figure BDA0001810470700000074
s3 retraining the network model: and (5) retraining the deep neural network quantified in the step S2. Inputting training images with class labels, obtaining cross entropy loss of the depth network in the forward process of training, and updating weight parameters in the network according to the cross entropy loss in the backward process.
In step S3, after the weights in the deep neural network are quantized in step S2, the deep neural network is retrained, which is divided into two procedures: forward propagation and backward propagation. In the forward propagation process, training data is input, and the cross entropy loss of the network is obtained in the deep neural network, which is defined as follows:
Figure BDA0001810470700000081
wherein the content of the first and second substances,
Figure BDA0001810470700000082
is a loss of the network and is,
Figure BDA0001810470700000083
is L2Norm regularization term, the invention adopts L2A regularization term, λ being a coefficient of the regularization term，
Figure BDA0001810470700000084
Is the network weight after the model compression,
Figure BDA0001810470700000085
is the total loss function. In the back propagation process, the residual error of the network is transmitted from the next layer to the front layer, the weight of the network is updated according to the gradient calculated by the residual error, and the updating mode is as follows:
Figure BDA0001810470700000086
wherein the content of the first and second substances,
Figure BDA00018104707000000811
is the updated weight, gamma is the learning rate,
Figure BDA0001810470700000087
is the gradient of the loss function over the network weights. For the quantization model, the derivation of the indicator function i (w) results in a gradient of 0, and the parameters cannot be updated. In the process of inverse derivation, the weights in the model can be processed in this way
Figure BDA0001810470700000088
The treatment method comprises the following steps:
Figure BDA0001810470700000089
therefore, in the actual back propagation process, the weight update mode is as follows:
Figure BDA00018104707000000810
s4 iteration and termination: the updating of the weight in step S3 destroys the original quantization, so that steps S2 and S3 are iteratively performed, and when the deep neural network model converges or reaches the set training times, the iteration is terminated to obtain the final quantization compression model.
S5 image object classification: the unclassified images are sent to the quantization compression model obtained in step S4 for prediction, and the images are classified according to the prediction result.
The image classification method based on the mobile terminal quantizes the weight in the deep neural network model into an index of 2, and can effectively improve the shift operation of the embedded system. Different from the existing method adopting static quantization coding, the method provided by the invention can be used for dynamically updating the quantization codebook in the model training process, effectively reducing the quantization error and improving the prediction performance of the model and the operation efficiency in the moving state.
Example (c):
this example is to propose an image classification device based on a mobile terminal, which includes three modules: an image reading module, an image classification module and an image sorting module, as shown in fig. 3.
This example was tested on the standard data set CIFAR-10. CIFAR-10 is an image classification dataset comprising 10 classes, respectively: airplanes, cars, birds, cats, deer, dogs, frogs, horses, boats, and trucks. All images are three-channel color images of size 32 × 32, containing 60000 pictures, of which the training set is 50000 and the validation set is 10000. The deep neural network employed in the experiment of this example was the residual network ResNet. The method comprises the following specific steps:
1. image reading
The test data were read one by one and the size was scaled to 32 x 32.
2. Image classification
And predicting the read image by using a dynamic quantized compressed depth neural network.
The training process of the compressed deep neural network comprises the following steps: performing enhancement preprocessing on the training data, namely filling images with 0 expansion of 36 × 36 on the original 32 × 32 image boundary, randomly cutting the images into 32 × 32 images, and then randomly turning the images left and right; the model is dynamically quantized and coded until the model converges, and in the training process, the model is dynamically quantized and coded80000 rounds of iteration are carried out, data fed into the network for one batch in each round is 128, the initial learning rate is 0.1, when training reaches 40000 times, the learning rate is 0.01, and after 60000 times, the learning rate is 0.001, and L is used in training2The coefficient of the regularization term is set to 0.001.
The prediction results of image classification are shown in table 1, and the cases of introducing 0 and not introducing 0 into the codebook are compared respectively.
Table 1 ResNet at different depths introduces 0 and no 0 effect on quantization in the codebook at different bit widths. The accuracy of the pre-trained 32-bit wide model on the verification set under ResNet-20, ResNet-32, ResNet-44 and ResNet-56 is 0.9212, 0.9246, 0.9332 and 0.9323 in sequence, and the fifth column and the seventh column in the table represent the accuracy of the quantized model on the verification set minus the accuracy of the pre-trained 32-bit wide model.
Figure BDA0001810470700000101
TABLE 1
As can be seen from Table 1, the method provided by the invention can effectively compress the deep neural network model to a very high multiple, the model can ensure higher performance, and even when the original model is compressed to 10.67 times, the performance of the model is only slightly reduced.
3. Image arrangement
And arranging the pictures into folders of corresponding categories according to the prediction result.
The deep neural network compression method has an obvious compression effect on the deep neural network, can reduce the consumption of a large deep neural network model on storage resources and calculation resources, promotes the deployment of the deep neural network on resource-limited mobile terminals such as smart phones and the like, and has extremely strong practical application.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (1)

1. An image classification method based on a mobile terminal is characterized by comprising an off-line training stage and an on-line classification stage:
the off-line training stage specifically comprises:
s1 formulating a codebook:
acquiring the maximum value of the weight absolute value in each layer of the deep neural network model by adopting a deep neural network model, and quantizing the maximum value into an exponential form with the base number of 2, thereby obtaining the quantization upper limit of the codebook; determining a codebook quantized by a current model under the limit of an upper quantization limit;
s2 quantization weight:
carrying out exponential quantization on the weight in the deep neural network model, and quantizing the weight into a value closest to a codebook;
s3 retraining the network model:
inputting a sample image, training a quantized deep neural network model, obtaining cross entropy loss of the deep network in the forward process of training, and updating weight parameters in the network according to the cross entropy loss in the backward process;
s4 iteration and termination:
iteratively executing steps S1, S2 and S3, dynamically updating the quantization codebook in the model training process until the deep neural network model converges or reaches the set training times, terminating iteration and obtaining a final classifier;
the online classification stage specifically comprises: sending the image to be classified into a classifier to obtain a classification result;
wherein the quantization upper limit of the codebook is represented as:
Figure FDA0002757041570000011
wherein n is2＝floor(log2(max(|WlI))) floor (. cndot.) is a downward rounding operation, wlFor the l-th layer weight of the deep neural network, max () represents the maximum value, and | represents the absolute value;
when quantized to b bits, the codebook is expressed as: pl＝{±2n},n∈[n11,n12]N ∈ Z, where l represents the l-th layer of the deep neural network, n11And n12Is two integers satisfying n11＜n12，n11＝n12-2b-1+1, Z represents a positive integer, or is represented by: pl＝{±2n,0},n∈[n21,n22]N ∈ Z, where l represents the l-th layer of the deep neural network, n21And n22Is two integers satisfying n21＜n22，n21＝n22-2b-2+1, Z represents a positive integer;
wherein, if the codebook is expressed as: pl＝{±2n},n∈[n11,n12]And n belongs to Z, quantizing each layer of the deep neural network as follows:
Figure FDA0002757041570000021
wherein the content of the first and second substances,
Figure FDA0002757041570000022
for quantized weights, 2kA quantized value that is the absolute value | w | of the weight w;
Figure FDA0002757041570000023
indicating function W ∈ Wl；
If the codebook is expressed as: pl＝{±2n,0},n∈[n21,n22]And n belongs to Z, quantizing each layer of the deep neural network as follows:
Figure FDA0002757041570000024
wherein the content of the first and second substances,
Figure FDA0002757041570000025
for quantized weights, 2kA quantized value that is the absolute value | w | of the weight w;
Figure FDA0002757041570000026
indicating function W ∈ Wl；
Obtaining cross entropy loss of the deep network in the forward process of training, wherein the cross entropy loss is expressed as:
Figure FDA0002757041570000031
wherein the content of the first and second substances,
Figure FDA0002757041570000032
is the loss of the network and is,
Figure FDA0002757041570000033
is a regularization term, adopts L2Norm regularization term, λ is the coefficient of the regularization term,
Figure FDA0002757041570000034
is the network weight after the model compression,
Figure FDA0002757041570000035
is the total loss function;
updating the weight parameters in the network according to the cross entropy loss in the reverse process:
Figure FDA0002757041570000036
wherein the content of the first and second substances,
Figure FDA0002757041570000037
is the weight of the network at the kth iteration, gamma is the learning rate,
Figure FDA0002757041570000038
is the gradient of the loss function over the network weights.
CN201811119618.4A 2018-09-25 2018-09-25 Image classification method based on mobile terminal Active CN109344893B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811119618.4A CN109344893B (en) 2018-09-25 2018-09-25 Image classification method based on mobile terminal

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811119618.4A CN109344893B (en) 2018-09-25 2018-09-25 Image classification method based on mobile terminal

Publications (2)

Publication Number Publication Date
CN109344893A CN109344893A (en) 2019-02-15
CN109344893B true CN109344893B (en) 2021-01-01

Family

ID=65306861

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811119618.4A Active CN109344893B (en) 2018-09-25 2018-09-25 Image classification method based on mobile terminal

Country Status (1)

Country Link
CN (1) CN109344893B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110245753A (en) * 2019-05-27 2019-09-17 东南大学 A kind of neural network compression method based on power exponent quantization
CN110414630A (en) * 2019-08-12 2019-11-05 上海商汤临港智能科技有限公司 The training method of neural network, the accelerated method of convolutional calculation, device and equipment
CN110782021B (en) * 2019-10-25 2023-07-14 浪潮电子信息产业股份有限公司 Image classification method, device, equipment and computer readable storage medium
CN113298224A (en) * 2020-02-24 2021-08-24 上海商汤智能科技有限公司 Retraining method of neural network model and related product
CN111582377A (en) * 2020-05-09 2020-08-25 济南浪潮高新科技投资发展有限公司 Edge end target detection method and system based on model compression
CN112668630B (en) * 2020-12-24 2022-04-29 华中师范大学 Lightweight image classification method, system and equipment based on model pruning
CN114462592A (en) * 2021-12-24 2022-05-10 光子算数(北京)科技有限责任公司 Model training method and device, electronic equipment and computer readable storage medium

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7089178B2 (en) * 2002-04-30 2006-08-08 Qualcomm Inc. Multistream network feature processing for a distributed speech recognition system
CN105590116B (en) * 2015-12-18 2019-05-14 华南理工大学 A kind of birds image-recognizing method based on head piece alignment
US10218976B2 (en) * 2016-03-02 2019-02-26 MatrixView, Inc. Quantization matrices for compression of video
CN106203624B (en) * 2016-06-23 2019-06-21 上海交通大学 Vector Quantization and method based on deep neural network
CN106713929B (en) * 2017-02-16 2019-06-28 清华大学深圳研究生院 A kind of video inter-prediction Enhancement Method based on deep neural network
CN107239793B (en) * 2017-05-17 2020-01-17 清华大学 Multi-quantization depth binary feature learning method and device
CN107423814A (en) * 2017-07-31 2017-12-01 南昌航空大学 A kind of method that dynamic network model is established using depth convolutional neural networks
CN108229681A (en) * 2017-12-28 2018-06-29 郑州云海信息技术有限公司 A kind of neural network model compression method, system, device and readable storage medium storing program for executing

Also Published As

Publication number Publication date
CN109344893A (en) 2019-02-15

Similar Documents

Publication Publication Date Title
CN109344893B (en) Image classification method based on mobile terminal
CN112101190B (en) Remote sensing image classification method, storage medium and computing device
EP3735658A1 (en) Generating a compressed representation of a neural network with proficient inference speed and power consumption
US11604960B2 (en) Differential bit width neural architecture search
CN110119745B (en) Compression method, compression device, computer equipment and storage medium of deep learning model
WO2019155064A1 (en) Data compression using jointly trained encoder, decoder, and prior neural networks
CN106778910B (en) Deep learning system and method based on local training
CN113850272A (en) Local differential privacy-based federal learning image classification method
CN109377532B (en) Image processing method and device based on neural network
CN113869420B (en) Text recommendation method and related equipment based on contrast learning
CN111008694A (en) No-data model quantization compression method based on deep convolution countermeasure generation network
WO2020207410A1 (en) Data compression method, electronic device, and storage medium
CN112200296A (en) Network model quantification method and device, storage medium and electronic equipment
CN113947136A (en) Image compression and classification method and device and electronic equipment
CN111582284B (en) Privacy protection method and device for image recognition and electronic equipment
CN113743277A (en) Method, system, equipment and storage medium for short video frequency classification
KR102305981B1 (en) Method for Training to Compress Neural Network and Method for Using Compressed Neural Network
CN116797850A (en) Class increment image classification method based on knowledge distillation and consistency regularization
CN116468947A (en) Cutter image recognition method, cutter image recognition device, computer equipment and storage medium
CN112200275B (en) Artificial neural network quantification method and device
CN112070211B (en) Image recognition method based on computing unloading mechanism
CN114677535A (en) Training method of domain-adaptive image classification network, image classification method and device
CN113438482A (en) Region of interest based video coding
Benbarrad et al. Impact of standard image compression on the performance of image classification with deep learning
CN113221560B (en) Personality trait and emotion prediction method, personality trait and emotion prediction device, computer device, and medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant